Skip to main content
Version: V1.0.0

seekdb MLDR benchmark report

This report summarizes the test plan, test environment, and benchmark results for the seekdb MLDR evaluation.

Test plan

  • Two machines are used:
    • One machine runs MLDR.
    • The other machine runs seekdb (4C8G). The log disk, clog disk, and data disk are placed on three separate disks, with performance level PL1.
  • Dataset size: 200,000 records.
  • The MLDR dataset is used to evaluate seekdb retrieval quality.

For step-by-step instructions, see seekdb MLDR benchmark.

Test environment

This benchmark was run on Alibaba Cloud ECS. The seekdb instance size is 4C8G.

Hardware

ServiceECS instance typeCountvCPUMemory
seekdbecs.c8i.xlarge14C8G

Software

ServiceVersion
seekdbseekdb 1.0.1
OSCentOS Linux 7 (Core)

Results

All MLDR results below are best-case response times after warm-up under a single-threaded setup. Results are returned with limit 1000.

BackendQuery type
seekdbdense+bm25:
  • ndcg@10 = 0.6329
  • recall@10 = 0.7425
  • rt = 150.39 ms